Closed OttCS closed 3 months ago
tested on Fedora 40:
cd stable-diffusion-webui
// remove previous attemps of installation
rm -rf venv
// edit webui-user.sh
// uncomment python_cmd
python_cmd="python3.10"
// uncomment TORCH_COMMAND
export TORCH_COMMAND="pip install torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm6.0"
sudo dnf install python3.10 gperftools-libs libglvnd-glx rocm-opencl
./webui.sh
Fedora 40 ships rocm6 while stable-diffusion-webui "officially" supports rocm 5.6 and 5.7. Prepare for a rough flight (applies in general with AMDGPU in compute space).
EDIT: don't bother with rocm 6.0 will crash every other generation. Using a container with Fedora 38 (which ships with rocm 5.5) is more stable in my experience.
After disabling AMD Integrated GPU, I only did:
$ rm -rf venv
$ sudo dnf install python3.10 gperftools-libs libglvnd-glx rocm-opencl
$ ./webui.sh
And it auto-fetched a working version of torch, no TORCH_COMMAND change necessary.
While this seems to fix the issue for Nobara (39), shouldn't the setup include changing the commented python version if it detects a bad one also installed?
I am having this exact issue, it seems we have a version of Python, Pytorch and CUDA that is too new and isnt properly recognized by WebUI. I am running a fully up to date Nobara (39) release and mirror the initial Nvidia related issue.
I am having this exact issue, it seems we have a version of Python, Pytorch and CUDA that is too new and isnt properly recognized by WebUI. I am running a fully up to date Nobara (39) release and mirror the initial Nvidia related issue.
Fedora includes the last 4 Python releases. While the latest is named just python
the previous include their version in their name python3.9
, python3.10
and so on.
Try:
sudo dnf install python3.10
// edit webui-user.sh of stable-diffusion-webui
// uncomment python_cmd
python_cmd="python3.10"
(delete the .venv
inside stable-diffusion-webui if your previous installation failed to avoid corrupted environments)
I am having this exact issue, it seems we have a version of Python, Pytorch and CUDA that is too new and isnt properly recognized by WebUI. I am running a fully up to date Nobara (39) release and mirror the initial Nvidia related issue.
Fedora includes the last 4 Python releases. While the latest is named just
python
the previous include their version in their namepython3.9
,python3.10
and so on.Try:
sudo dnf install python3.10 // edit webui-user.sh of stable-diffusion-webui // uncomment python_cmd python_cmd="python3.10"
(delete the
.venv
inside stable-diffusion-webui if your previous installation failed to avoid corrupted environments)
Did as instructed, and it installed! I will come back incase it finds a reason to not function
Edit: Appears to work, generates images and everything!
Judging by how reliable these fixes are, I made a small (unofficial) guide for how to install this on Linux, hopefully someone can tell me how to contribute to this wiki so the information is easier to find.
https://github.com/OttCS/OttCS/blob/main/blog/2024/5/linux-automatic1111-sd.md
Closing the issue, as the linux install instructions have been updated
Checklist
What happened?
The provided installation instructions for Linux don't seem to be working, at least for Nobara (which is Fedora based). The error printout is below, but of note:
sudo dnf install python3
currently installs version 3.12.2 which doesn't support torch.sudo dnf install python3.10
), the ./webui.sh launch doesn't detect this.sudo pacman -S pyenv
orsudo pacman -Sy pyenv
)Something may be configured wrong on Nobara's part, but I'm willing to help with making the guides for webui on Linux more reliable.
Steps to reproduce the problem
What should have happened?
WebUI should be able to run on Nobara (and actually used to before some time in the past month)
What browsers do you use to access the UI ?
Other
Sysinfo
dump-sysinfo.txt
Console logs
Additional information
Browser selection is should not matter, as WebUI is failing before opening.